Surface Area Heuristic for Bounding Value Hierarchy structures

[Walther]

Implementing the Surface Area Heuristic algorithm for Bounding Volume Hierarchy tree construction.

Based on these two blog articles:

• https://jacco.ompf2.com/2022/04/18/how-to-build-a-bvh-part-2-faster-rays/
• https://jacco.ompf2.com/2022/04/21/how-to-build-a-bvh-part-3-quick-builds/

The new SAH algorithm results in much more efficient BVH trees. This affects especially scenes with a larger amount of primitives.

hyperfine -L bvh lam,sah "just cli render -w 1920 -h 1080 --sampler blue --samples 8 --bvh {bvh} -o temp.png ~/sponza.json"
Benchmark 1: just cli render -w 1920 -h 1080 --sampler blue --samples 8 --bvh lam -o temp.png ~/sponza.json
  Time (mean ± σ):     52.065 s ±  0.318 s    [User: 387.275 s, System: 0.751 s]
  Range (min … max):   51.794 s … 52.789 s    10 runs
 
Benchmark 2: just cli render -w 1920 -h 1080 --sampler blue --samples 8 --bvh sah -o temp.png ~/sponza.json
  Time (mean ± σ):     20.147 s ±  0.050 s    [User: 141.036 s, System: 0.381 s]
  Range (min … max):   20.090 s … 20.255 s    10 runs
 
Summary
  just cli render -w 1920 -h 1080 --sampler blue --samples 8 --bvh sah -o temp.png ~/sponza.json ran
    2.58 ± 0.02 times faster than just cli render -w 1920 -h 1080 --sampler blue --samples 8 --bvh lam -o temp.png ~/sponza.json
hyperfine -L bvh lam,sah   5283.53s user 11.72s system 733% cpu 12:02.34 total

There's also a couple of new debug rendering modes, for visualizing the BVH node test count and primitive test count needed per pixel.

Before, with Longest Axis Midpoint BVH construction

After, with Surface Area Heuristic BVH construction

Colorization key